A friend pointed me to a fun puzzle. There’s a fairly-straightforward “standard” solution, but I had some extra fun finding increasingly weird ones.

This post is really about abstract algebra, but I’m going to try to keep the prerequisites low by spelling everything out. Target audience: people who like abstract algebra, mathy puzzles, or both.

## The puzzle

Here’s the puzzle: https://www.futilitycloset.com/2021/08/27/50-for-all/

A puzzle by Ben H., a systems engineer at the National Security Agency, from the agency’s August 2016 Puzzle Periodical:

At a work picnic, Todd announces a challenge to his coworkers. Bruce and Ava are selected to play first. Todd places $100 on a table and explains the game. Bruce and Ava will each draw a random card from a standard 52-card deck. Each will hold that card to his/her forehead for the other person to see, but neither can see his/her own card. The players may not communicate in any way. Bruce and Ava will each write down a guess for the color of his/her own card, i.e. red or black. If either one of them guesses correctly, they both win $50. If they are both incorrect, they lose. He gives Bruce and Ava five minutes to devise a strategy beforehand by which they can guarantee that they each walk away with the $50.

Bruce and Ava complete their game and Todd announces the second level of the game. He places $200 on the table. He tells four of his coworkers — Emily, Charles, Doug and Fran — that they will play the same game, except this time guessing the suit of their own card, i.e. clubs, hearts, diamonds or spades. Again, Todd has the four players draw cards and place them on their foreheads so that each player can see the other three players’ cards, but not his/her own. Each player writes down a guess for the suit of his/her own card. If at least one of them guesses correctly, they each win $50. There is no communication while the game is in progress, but they have five minutes to devise a strategy beforehand by which they can be guaranteed to walk away with $50 each.

For each level of play — 2 players or 4 players — how can the players ensure that someone in the group always guesses correctly?

I strongly encourage you to solve this puzzle yourself before reading on. I was tempted to read the answer, but had more fun figuring it out (I went down a blind alley at first).

## The standard solution (for N players)

We can generalize this puzzle to the case with N different “colors” or “suits” of cards and N players.

Here’s the standard solution: (For a less mathy version, see the Futility Closet link.)

- Assign the players numbers from 0 to N-1.
- Assign each suit a different number from 0 to N-1.
- Consider the sum of all of the players’ cards’ suit-numbers (mod N). This is also a number between 0 and N-1.
- Strategy: Player
**k**is assigned the hypothesis “the sum of the suits is**k**“. They guess that their own card has whatever suit is needed to make the sum come out to**k**. - Concretely, each player takes their own player-number, subtracts each of the other players’ suit-numbers (mod N), and guesses the resulting suit.
- One player will be assigned the correct hypothesis, so they will guess their own suit correctly.

More generally, suppose we have a function **W **from N card suits to a number mod N.

(The standard solution (for N=4) is to let **W**(a, b, c, d) = (a + b + c + d) % 4, where “%” is the mod operator.)

Each player guesses a different value for **W**. For example, taking N=4 again, the third player always guesses that **W**=2. In order for the strategy to work, the equation **W**(a, b, c, d) = 2 needs to have a unique solution for c in terms of a, b, d. (And similarly, **W**(a, b, c, d) = 0 needs to be solvable for a, and so forth.)

This is easy to verify for the standard solution, but it’s not immediately obvious how to construct other functions with this property.

## Rows by any other name

What if we want a different strategy? Well, to start with, we can give the players different numbers — maybe Doug is 3 and Fran is 2 instead of vice-versa. Or we can give the suits different numbers. (If we do both, some combinations of choices will give the same strategy, though.)

There are some easy and dumb ways to go further. For one thing, the assignment of numbers to suits can be different depending on whose card it is (but **not** depending on who’s looking at it!).

Tricks like this give a pretty large number of essentially-the-same solutions, which I’m not really interested in counting.

## More ways to add: abelian groups

Addition mod N is a way to “add up” a bunch of numbers in the set {0, 1, …, N-1} and also lets you subtract them. There are other ways to do this: an example for N=4 is to write numbers in binary, {00, 01, 10, 11}, and separately add the bits mod 2. (You can think of this as bitwise-XORing the numbers if you like to code in C.)

So there’s another solution: **W**(a, b, c, d) = a XOR b XOR c XOR d.

(Exercise: Show that this works if and only if N is a power of 2.)

What can we do for arbitrary N?

Well, we can generalize the two-bit strategy: if N = X * Y, we can map our N numbers to pairs (x, y), where 0 <= x < X and 0 <= y < Y. Then we can add two pairs by adding their parts separately, mod X and mod Y respectively. And we can get recursive, by breaking down X = Z * W….

In general, any way of adding and subtracting things — such that the order you do it in doesn’t matter — is an abelian group. (I’m hiding some axioms in the words “adding and subtracting”. For example, adding something and then subtracting it shouldn’t do anything.)

There’s a cute theorem that says that any finite abelian group is of the form above: it’s a direct sum of cyclic groups. (A “cyclic group” is addition mod N for some N, and a “direct sum” is the pairing construction above. Confusingly, “direct product” and “direct sum” mean the same thing in this context.)

So we’re done, right?

## Non-commutative groups

Not all groups are abelian — sometimes the order matters. Also, not all groups are groups of numbers; sometimes they’re groups of other things.

Here’s a cute example: take the N=24 case. Instead of numbers from 0 to 23, assign each player a different way to permute four things. For example, maybe Doug is assigned “swap the first two things and leave the other two in place”. There are 4! = 24 ways to permute four things, so this works. Also, assign each suit a different 4-permutation as well.

Now, when you go to “add up” some suit and player permutations, do it by setting out four objects, applying each permutation one after another, and then seeing what overall permutation you end up with. In other words, **W**(a, b, c, d) = a∘b∘c∘d, where (a, b, c, d) are four permutations and “x∘y” represents composition of permutations (doing x and then doing y).

This time, it matters what order you do it in (e.g. whether you do Doug’s permutation before or after Fran’s). (Exercise: Find two permutations that don’t commute: a∘b ≠ b∘a.)

The strategy is still the same: “adding up” everyone else’s cards and “subtracting” that from your own number. The only difference is that it’s kind of annoying to work out the order to do that in, and the subtraction has to happen somewhere in the middle sometimes. (Exercise: Fill in the details here, and explain what “subtracting” a permutation means.)

In addition to modular arithmetic and permutation groups, there are tons of other places groups show up — symmetries of geometric objects are one of my favorites. This post is already too long, so I’ll save the group propaganda for the future.

Is there a cute theorem that tells you what all the possible finite groups are?

No, there’s a fairly messy theorem with an *enormous* proof. It took hundreds (thousands?) of mathematician-years to prove it, and the result was so long (sprawling across countless papers in math journals) that a few brave souls have been working on a condensed “second-generation” version. The first volume of the second-generation proof was published in 1994; the ninth volume came out this year. There will probably be four more volumes. The group classification megaproject is an incredible achievement, and one of the worst theorems I’ve ever heard of.

Worse yet, it only classifies the finite *simple* groups. I won’t define “simple” here, but the upshot is:

- Every finite group can be broken down into a bunch of simple groups
- …But doing the reverse is really hard; there’s no known way to quickly figure out all the ways to build up a bunch of simple groups into a bigger group.

So it’s not really clear, philosophically, that what I’ve given you in this section is a “solution” to the puzzle at all. What I’ve given you is a machine for turning finite groups into solutions. You also need some feedstock for the machine: here‘s all the groups for N<32.

But there’s no known systematic way to find all the groups of order N, for arbitrary N. We don’t even know how many groups of order 2048 there are.

## Mixing groups together

Here’s another thing you can do: add up the four suits (a, b, c, d) as follows:

**W**(a, b, c, d) = (a ∘ b) + (c # d)

…where ∘, +, and # are the group operations for three *different* groups.

So, for example, you can add the first two and last two numbers mod 4, then convert the sums to binary and add them (mod 2, mod 2).

(Exercise: Show that this is still a puzzle solution (i.e. you can still solve for a, b, c, or d.))

When you need to “translate” between groups (e.g. turn a permutation-of-**n**-things into a number-mod-**n!**), you can do it totally arbitrarily! Just make sure the translation you use is a one-to-one mapping. (Exercise: show that this is okay.)

This gives a huge explosion of strategies. Also, now the overall operation isn’t even associative, so you have to be careful and not do **W** = a ∘ ((b + c) # d) by mistake — that’s a totally different solution.

## Quasigroups and Latin squares

Fun fact: despite all the terrible solutions above, none of them give interestingly-different answers to the cases where N is prime, such as N=2, 3, or 5. The only variety for those cases comes from the relabeling tricks: assigning different card-number mappings per player, and “translating” intermediate values in the computation via arbitrary permutations. That’s because the only groups of prime order are cyclic groups, which give you the standard solution to the puzzle.

It’s time to rectify that. (There’s a pun here somewhere — “rectify”, Latin, *rectus*, square, rectangle… oh well.)

Groups have to be associative, so that (a∘b)∘c = a∘(b∘c). As we just saw, solutions to this puzzle don’t need to be associative, so there’s no need to limit ourselves to binary operations that come from groups.

There’s two perspectives here:

(1): A group without the requirement of associativity is a quasigroup. Group theory is a beautiful subject; quasigroup theory is an ugly one.

(2): The multiplication table of a quasigroup is a Latin square, which is basically a solved Sudoku without the smaller 3×3 boxes.

There’s no hope of enumerating these quickly or easily. By relaxing the amount of structure of our solutions too far, we have passed beyond the event horizon of classifiability. There is no theorem for these, nice or otherwise. They have some cool connections, though, like the design of statistical experiments. And they look nice in stained glass.

## Latin (hyper)cubes

For a binary operator, the multiplication table is a square. But there’s no need to restrict ourselves to combinations of binary operators.

A Latin cube (or hypercube) is like a Latin square, but in more dimensions.

With three Latin squares X, Y, Z, we can write equations like

**W**(a, b, c, d) = X[Y[a, b], Z[c, d]],

where X[i, j] is the entry in the i-th row and j-th column of X. This should look pretty familiar; it’s the same as the expression (a Y b) X (c Z d) for the corresponding quasigroup operations X, Y, Z.

With a Latin cube C and a square S, we instead have things like

**W**(a, b, c, d) = C[a, S[b, c], d]

You could interpret the Latin cube as the multiplication table for a ternary operator in some sort of awful algebraic structure, but why make life hard for yourself?

## Horrible lopsided solutions

A Latin N-cube is *almost* the last word on solutions to this puzzle, since it’s equivalent to a function **W **such that

**W**(a, b, c….) = w

can be solved uniquely for a, or b, or c, etc.

But this is slightly stronger than what we asked for, which was:

In order for the strategy to work, the equation

W(a, b, c, d) = 2 needs to have a unique solution for c in terms of a, b, d. (And similarly,W(a, b, c, d) = 0 needs to be solvable for a, and so forth.)

We only need the equation to be solvable for the* w*-th argument (counting from zero).

But are there any solutions like this that aren’t full Latin hypercubes?

Yup. It’s not too hard to find one with a backtracking search; I did N=3 by hand in a few minutes.

## Ne plus ultra

Any solution to the puzzle has to be of this form.

(Exercise: Show that every solution has exactly one winner, no matter what the suits are. **W** specifies this winner in terms of those suits. Show that every strategy is equivalent to “assume the winner is going to be you, then solve the equation * W(a, b, c…) = You* to find your suit”.)

But by the same token, this form is not very helpful; a “generalized Latin hypercube” **W **is just an arbitrary array of N^N values, with the constraint that it must solve this puzzle.

The solutions in this post are on a spectrum, with the standard solution being the simplest and easiest to execute, and later solutions being increasingly weird and also increasingly hard to construct explicitly. I think the most interesting ones are probably the non-abelian groups, since they have the most fun math involved (pick your favorite Platonic solid; its group of symmetries specifies a puzzle solution for 24, 48, or 120 players) but your mileage may vary.

It’s kind of like this:

On the left, the standard solution. In the middle, group solutions. On the right, Latin hypercube solutions.

Ponder this mystery.